train_config.yaml

data: {train_data: "/home/coherent_pro_full", eval_data: "", shuffle: true}
moshi_paths: {hf_repo_id: "kyutai/moshiko-pytorch-bf16"}
full_finetuning: false
lora: {enable: true, rank: 64, scaling: 2., ft_embed: false}
first_codebook_weight_multiplier: 100.
text_padding_weight: .5
duration_sec: 20
batch_size: 1
max_steps: 1500
gradient_checkpointing: true
optim: {lr: 2.e-5, weight_decay: 0.1, pct_start: 0.05}
seed: 0
log_freq: 50
eval_freq: 20000
do_eval: false
do_ckpt: true
ckpt_freq: 1500
save_adapters: true
run_dir: "/home/moshi_run_pro_full"